Chris Rackauckas: fine-tuned local models run robot heads on Orin 64GB, Opus writes the scripts
ChrisRackauckas · x · 2026-10-18
Chris Rackauckas (SciML/Julia lead) shares hands-on practice: his team fine-tunes local models to run on Orin 64GB heads of autonomous robots serving beer and managing a gaming cafe on an island, with much lower latency via onboard tuned models.
His division of labor from earlier posts:
- Open models win on routine maintenance tasks — Opus wrote scripts for mass rebasing, CI regression checks on current PRs, and auto-merge risk assessments with written criteria; then local models handle it 100% and cheaper;
- For models that can find deep code defects, he's scanning everything on Fireworks — so far nothing besides Sol and Opus looks promising.
Related event: MIT's Rackauckas Fine-Tunes Local Models for Robots Now Working Real Jobs(2 posts)→
More from coding & agent
- Meta paper introduces "agent plasticity": best-performing agent isn't the most efficient learner — rohanpaul_ai · 2026-10-18
- Dev ships a bushfire prediction platform in a weekend with frontier models — CtrlAltDwayne · 2026-10-18
- Developer launches Is it vibe coded?, a site that detects AI-built websites by design fingerprints — sahilypatel · 2026-10-18
- 10 open-source GitHub repos that let AI agents control your phone, from Open-AutoGLM to mobile-use — victor_explore · 2026-10-18
- Indie dev's 7 AI projects in a year: hiring agent flopped, WeChat automation custom gig made ~$14K — sujingshen · 2026-10-18
- Dev with 108K followers: Codex is now slower than typing code by hand — lxfater · 2026-10-18